Built a small open-source tool I wanted to share: FlowProof, an MCP server that lets an AI assistant run Nextflow pipelines and returns a Workflow Run RO-Crate for each run (container digests, tool versions, SHA-256 of every input and output), so AI-driven pipeline runs are reproducible and verifiable rather than a black box. It sits above the engine (in the spirit of nf-prov, but the provenance is retrievable through the same interface that launched the run). Runs locally (uvx flowproof-mcp, your data stays on your machine) or hosted. Repo: github.com/ajibadedapo/flowproof-mcp It's early, and feedback is very welcome, especially on the provenance model and which pipeline you'd want wired next.

Does FlowProof’s output collection handle an

s3://

--outdir, or do

output_globs

only resolve against the local

run_dir

? Running against a cloud executor, I think that’d give an empty results manifest / no checksums. Is staging results back locally the intended pattern, or would S3 support be PR-able?

That's exactly how it behaves today. Output collection is local-only so it isn't a user-supplied s3:// outdir to begin with. Native S3 support is very PR-able and I'd welcome it.